138ce3211762d27055577ae7b31623e5

nano.ipynb

notebook.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/HudsonGraeme/138ce3211762d27055577ae7b31623e5/notebook.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VOhZf5ompktU"
      },
      "source": [
        "### Install required dependencies"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "id": "m-28wslFI2wX",
        "outputId": "a7598887-4d4e-4cf4-eee8-58cead4ad631",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current\n",
            "                                 Dload  Upload   Total   Spent    Left  Speed\n",
            "100  5718  100  5718    0     0  27411      0 --:--:-- --:--:-- --:--:-- 27490\n",
            "No version tags provided, installing the latest ezkl version\n",
            "Removing old ezkl binary if it exists\n",
            "Downloading package\n",
            "  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current\n",
            "                                 Dload  Upload   Total   Spent    Left  Speed\n",
            "  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0\n",
            "100 11.2M  100 11.2M    0     0  12.0M      0 --:--:-- --:--:-- --:--:-- 16.2M\n",
            "Unpacking package\n",
            "Cleaning up\n",
            "\n",
            "Successfully downloaded ezkl at /root/.ezkl\n",
            "We detected that your preferred shell is bash and added ezkl to PATH. Run 'source /root/.bashrc' or start a new terminal session to use ezkl.\n",
            "\u001b[33mWARNING: Skipping tensorflow as it is not installed.\u001b[0m\u001b[33m\n",
            "\u001b[0mRequirement already satisfied: dataclasses in /usr/local/lib/python3.10/dist-packages (0.6)\n",
            "Requirement already satisfied: matplotlib in /usr/local/lib/python3.10/dist-packages (3.7.1)\n",
            "Requirement already satisfied: contourpy>=1.0.1 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (1.2.0)\n",
            "Requirement already satisfied: cycler>=0.10 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (0.12.1)\n",
            "Requirement already satisfied: fonttools>=4.22.0 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (4.46.0)\n",
            "Requirement already satisfied: kiwisolver>=1.0.1 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (1.4.5)\n",
            "Requirement already satisfied: numpy>=1.20 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (1.23.5)\n",
            "Requirement already satisfied: packaging>=20.0 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (23.2)\n",
            "Requirement already satisfied: pillow>=6.2.0 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (9.4.0)\n",
            "Requirement already satisfied: pyparsing>=2.3.1 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (3.1.1)\n",
            "Requirement already satisfied: python-dateutil>=2.7 in /usr/local/lib/python3.10/dist-packages (from matplotlib) (2.8.2)\n",
            "Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.10/dist-packages (from python-dateutil>=2.7->matplotlib) (1.16.0)\n",
            "Requirement already satisfied: torch in /usr/local/lib/python3.10/dist-packages (2.1.0+cu121)\n",
            "Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from torch) (3.13.1)\n",
            "Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch) (4.9.0)\n",
            "Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch) (1.12)\n",
            "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch) (3.2.1)\n",
            "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from torch) (3.1.2)\n",
            "Requirement already satisfied: fsspec in /usr/local/lib/python3.10/dist-packages (from torch) (2023.6.0)\n",
            "Requirement already satisfied: triton==2.1.0 in /usr/local/lib/python3.10/dist-packages (from torch) (2.1.0)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->torch) (2.1.3)\n",
            "Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch) (1.3.0)\n",
            "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (1.23.5)\n",
            "Requirement already satisfied: requests in /usr/local/lib/python3.10/dist-packages (2.31.0)\n",
            "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests) (3.3.2)\n",
            "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests) (3.6)\n",
            "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests) (2.0.7)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests) (2023.11.17)\n",
            "Requirement already satisfied: onnxruntime in /usr/local/lib/python3.10/dist-packages (1.16.3)\n",
            "Requirement already satisfied: coloredlogs in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (15.0.1)\n",
            "Requirement already satisfied: flatbuffers in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (23.5.26)\n",
            "Requirement already satisfied: numpy>=1.21.6 in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (1.23.5)\n",
            "Requirement already satisfied: packaging in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (23.2)\n",
            "Requirement already satisfied: protobuf in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (3.20.3)\n",
            "Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from onnxruntime) (1.12)\n",
            "Requirement already satisfied: humanfriendly>=9.1 in /usr/local/lib/python3.10/dist-packages (from coloredlogs->onnxruntime) (10.0)\n",
            "Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->onnxruntime) (1.3.0)\n",
            "Requirement already satisfied: onnx in /usr/local/lib/python3.10/dist-packages (1.15.0)\n",
            "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (from onnx) (1.23.5)\n",
            "Requirement already satisfied: protobuf>=3.20.2 in /usr/local/lib/python3.10/dist-packages (from onnx) (3.20.3)\n",
            "Requirement already satisfied: torchvision in /usr/local/lib/python3.10/dist-packages (0.16.0+cu121)\n",
            "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (from torchvision) (1.23.5)\n",
            "Requirement already satisfied: requests in /usr/local/lib/python3.10/dist-packages (from torchvision) (2.31.0)\n",
            "Requirement already satisfied: torch==2.1.0 in /usr/local/lib/python3.10/dist-packages (from torchvision) (2.1.0+cu121)\n",
            "Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /usr/local/lib/python3.10/dist-packages (from torchvision) (9.4.0)\n",
            "Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (3.13.1)\n",
            "Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (4.9.0)\n",
            "Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (1.12)\n",
            "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (3.2.1)\n",
            "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (3.1.2)\n",
            "Requirement already satisfied: fsspec in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (2023.6.0)\n",
            "Requirement already satisfied: triton==2.1.0 in /usr/local/lib/python3.10/dist-packages (from torch==2.1.0->torchvision) (2.1.0)\n",
            "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests->torchvision) (3.3.2)\n",
            "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests->torchvision) (3.6)\n",
            "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests->torchvision) (2.0.7)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests->torchvision) (2023.11.17)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->torch==2.1.0->torchvision) (2.1.3)\n",
            "Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch==2.1.0->torchvision) (1.3.0)\n",
            "Found existing installation: typing_extensions 4.9.0\n",
            "Uninstalling typing_extensions-4.9.0:\n",
            "  Successfully uninstalled typing_extensions-4.9.0\n",
            "Collecting typing_extensions\n",
            "  Using cached typing_extensions-4.9.0-py3-none-any.whl (32 kB)\n",
            "Installing collected packages: typing_extensions\n",
            "\u001b[31mERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n",
            "tensorflow-probability 0.22.0 requires typing-extensions<4.6.0, but you have typing-extensions 4.9.0 which is incompatible.\u001b[0m\u001b[31m\n",
            "\u001b[0mSuccessfully installed typing_extensions-4.9.0\n"
          ]
        }
      ],
      "source": [
        "!curl https://raw.githubusercontent.com/zkonduit/ezkl/main/install_ezkl_cli.sh | bash\n",
        "!pip uninstall -y tensorflow\n",
        "!pip install dataclasses\n",
        "!pip install matplotlib\n",
        "!pip install torch\n",
        "!pip install numpy\n",
        "!pip install requests\n",
        "!pip install onnxruntime\n",
        "!pip install onnx\n",
        "!pip install torchvision\n",
        "!pip uninstall -y typing_extensions\n",
        "!pip install typing_extensions\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "P2iv9gZLpktV"
      },
      "source": [
        "### Model Definition"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "-61wEcCP1OZ5"
      },
      "outputs": [],
      "source": [
        "\"\"\"\n",
        "Reference: https://github.com/karpathy/nanoGPT\n",
        "\"\"\"\n",
        "\n",
        "import json\n",
        "import math\n",
        "import inspect\n",
        "from dataclasses import dataclass\n",
        "import torch\n",
        "import torch.nn as nn\n",
        "from torch.nn import functional as F\n",
        "import sys\n",
        "import os\n",
        "import onnxruntime as ort\n",
        "import numpy as np\n",
        "import subprocess\n",
        "import re\n",
        "\n",
        "\n",
        "def remove_non_ascii(s):\n",
        "    regex = re.compile(r\"\\x1b\\[([0-9]*;?[0-9]+)?[m|K|h]\")\n",
        "    return regex.sub(\"\", s)\n",
        "\n",
        "\n",
        "def new_gelu(x):\n",
        "    \"\"\"\n",
        "    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT).\n",
        "    Reference: Gaussian Error Linear Units (GELU) paper: https://arxiv.org/abs/1606.08415\n",
        "    \"\"\"\n",
        "    return (\n",
        "        0.5\n",
        "        * x\n",
        "        * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * x * x * x)))\n",
        "    )\n",
        "\n",
        "\n",
        "class LayerNorm(nn.Module):\n",
        "    \"\"\"LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False\"\"\"\n",
        "\n",
        "    def __init__(self, ndim, bias):\n",
        "        super().__init__()\n",
        "        self.weight = nn.Parameter(torch.ones(ndim))\n",
        "        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None\n",
        "\n",
        "    def forward(self, input):\n",
        "        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)\n",
        "\n",
        "\n",
        "class CausalSelfAttention(nn.Module):\n",
        "    def __init__(self, config):\n",
        "        super().__init__()\n",
        "        assert config.n_embd % config.n_head == 0\n",
        "        # key, query, value projections for all heads, but in a batch\n",
        "        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)\n",
        "        # output projection\n",
        "        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)\n",
        "        # regularization\n",
        "        self.attn_dropout = nn.Dropout(config.dropout)\n",
        "        self.resid_dropout = nn.Dropout(config.dropout)\n",
        "        self.n_head = config.n_head\n",
        "        self.n_embd = config.n_embd\n",
        "        self.dropout = config.dropout\n",
        "\n",
        "        # causal mask to ensure that attention is only applied to the left in the input sequence\n",
        "        self.register_buffer(\n",
        "            \"bias\",\n",
        "            torch.tril(torch.ones(config.block_size, config.block_size)).view(\n",
        "                1, 1, config.block_size, config.block_size\n",
        "            ),\n",
        "        )\n",
        "\n",
        "    def forward(self, x):\n",
        "        (\n",
        "            B,\n",
        "            T,\n",
        "            C,\n",
        "        ) = x.size()  # batch size, sequence length, embedding dimensionality (n_embd)\n",
        "        # calculate query, key, values for all heads in batch and move head forward to be the batch dim\n",
        "        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)\n",
        "        k = k.view(B, T, self.n_head, C // self.n_head).transpose(\n",
        "            1, 2\n",
        "        )  # (B, nh, T, hs)\n",
        "        q = q.view(B, T, self.n_head, C // self.n_head).transpose(\n",
        "            1, 2\n",
        "        )  # (B, nh, T, hs)\n",
        "        v = v.view(B, T, self.n_head, C // self.n_head).transpose(\n",
        "            1, 2\n",
        "        )  # (B, nh, T, hs)\n",
        "\n",
        "        # manual implementation of attention\n",
        "        # q shape:(B, nh, T, hs), k transpose shape (B, nh, hs, T) -> (B, nh, T, T)\n",
        "        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))\n",
        "        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float(-10))\n",
        "        att = F.softmax(att, dim=-1)\n",
        "        att = self.attn_dropout(att)\n",
        "        y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n",
        "        # re-assemble all head outputs side by side\n",
        "        y = y.transpose(1, 2).contiguous().view(B, T, C)\n",
        "\n",
        "        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)\n",
        "        # output projection\n",
        "        y = self.resid_dropout(self.c_proj(y))\n",
        "        return y\n",
        "\n",
        "\n",
        "class MLP(nn.Module):\n",
        "    def __init__(self, config):\n",
        "        super().__init__()\n",
        "        self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)\n",
        "        self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)\n",
        "        self.dropout = nn.Dropout(config.dropout)\n",
        "\n",
        "    def forward(self, x):\n",
        "        x = self.c_fc(x)\n",
        "        x = new_gelu(x)\n",
        "        x = self.c_proj(x)\n",
        "        x = self.dropout(x)\n",
        "        return x\n",
        "\n",
        "\n",
        "class Block(nn.Module):\n",
        "    def __init__(self, config):\n",
        "        super().__init__()\n",
        "        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)\n",
        "        self.attn = CausalSelfAttention(config)\n",
        "        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)\n",
        "        self.mlp = MLP(config)\n",
        "\n",
        "    def forward(self, x):\n",
        "        x = x + self.attn(self.ln_1(x))\n",
        "        x = x + self.mlp(self.ln_2(x))\n",
        "        return x\n",
        "\n",
        "\n",
        "@dataclass\n",
        "class GPTConfig:\n",
        "    block_size: int = 1024\n",
        "    # GPT-2 vocab_size of 50257, padded up to nearest multiple of 64 for efficiency\n",
        "    vocab_size: int = 50304\n",
        "    n_layer: int = 12\n",
        "    n_head: int = 12\n",
        "    n_embd: int = 768\n",
        "    dropout: float = 0.0\n",
        "    # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster\n",
        "    bias: bool = True\n",
        "\n",
        "\n",
        "class GPT(nn.Module):\n",
        "    def __init__(self, config, train_mode=False):\n",
        "        super().__init__()\n",
        "        assert config.vocab_size is not None\n",
        "        assert config.block_size is not None\n",
        "        self.config = config\n",
        "        self.train_mode = train_mode\n",
        "\n",
        "        self.transformer = nn.ModuleDict(\n",
        "            dict(\n",
        "                wte=nn.Embedding(config.vocab_size, config.n_embd),\n",
        "                wpe=nn.Embedding(config.block_size, config.n_embd),\n",
        "                drop=nn.Dropout(config.dropout),\n",
        "                h=nn.ModuleList([Block(config) for _ in range(config.n_layer)]),\n",
        "                ln_f=LayerNorm(config.n_embd, bias=config.bias),\n",
        "            )\n",
        "        )\n",
        "\n",
        "        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)\n",
        "\n",
        "        # weight-tying\n",
        "        # https://paperswithcode.com/method/weight-tying\n",
        "        self.transformer.wte.weight = self.lm_head.weight\n",
        "        self.block = Block(config)\n",
        "        # init all weights\n",
        "        self.apply(self._init_weights)\n",
        "        # apply special scaled init to the residual projections, per GPT-2 paper\n",
        "        for pn, p in self.named_parameters():\n",
        "            if pn.endswith(\"c_proj.weight\"):\n",
        "                torch.nn.init.normal_(\n",
        "                    p, mean=0.0, std=0.02 / math.sqrt(2 * config.n_layer)\n",
        "                )\n",
        "\n",
        "        # report number of parameters\n",
        "        print(\"number of parameters: %.2fM\" % (self.get_num_params() / 1e6,))\n",
        "\n",
        "    def get_num_params(self, non_embedding=True):\n",
        "        \"\"\"\n",
        "        Return the number of parameters in the model.\n",
        "        For non-embedding count (default), the position embeddings get subtracted.\n",
        "        The token embeddings would too, except due to the parameter sharing these\n",
        "        params are actually used as weights in the final layer, so we include them.\n",
        "        \"\"\"\n",
        "        n_params = sum(p.numel() for p in self.parameters())\n",
        "        if non_embedding:\n",
        "            n_params -= self.transformer.wpe.weight.numel()\n",
        "        return n_params\n",
        "\n",
        "    def _init_weights(self, module):\n",
        "        if isinstance(module, nn.Linear):\n",
        "            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)\n",
        "            if module.bias is not None:\n",
        "                torch.nn.init.zeros_(module.bias)\n",
        "        elif isinstance(module, nn.Embedding):\n",
        "            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)\n",
        "\n",
        "    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):\n",
        "        # start with all of the candidate parameters\n",
        "        param_dict = {pn: p for pn, p in self.named_parameters()}\n",
        "        # filter out those that do not require grad\n",
        "        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}\n",
        "        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.\n",
        "        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.\n",
        "        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]\n",
        "        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]\n",
        "        optim_groups = [\n",
        "            {\"params\": decay_params, \"weight_decay\": weight_decay},\n",
        "            {\"params\": nodecay_params, \"weight_decay\": 0.0},\n",
        "        ]\n",
        "        num_decay_params = sum(p.numel() for p in decay_params)\n",
        "        num_nodecay_params = sum(p.numel() for p in nodecay_params)\n",
        "        print(\n",
        "            f\"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters\"\n",
        "        )\n",
        "        print(\n",
        "            f\"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters\"\n",
        "        )\n",
        "        # Create AdamW optimizer and use the fused version if it is available\n",
        "        fused_available = \"fused\" in inspect.signature(torch.optim.AdamW).parameters\n",
        "        use_fused = fused_available and device_type == \"cuda\"\n",
        "        extra_args = dict(fused=True) if use_fused else dict()\n",
        "        optimizer = torch.optim.AdamW(\n",
        "            optim_groups, lr=learning_rate, betas=betas, **extra_args\n",
        "        )\n",
        "        print(f\"using fused AdamW: {use_fused}\")\n",
        "        return optimizer\n",
        "\n",
        "    def estimate_mfu(self, fwdbwd_per_iter, dt):\n",
        "        \"\"\"estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS\"\"\"\n",
        "        # first estimate the number of flops we do per iteration.\n",
        "        # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311\n",
        "        N = self.get_num_params()\n",
        "        cfg = self.config\n",
        "        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd // cfg.n_head, cfg.block_size\n",
        "        flops_per_token = 6 * N + 12 * L * H * Q * T\n",
        "        flops_per_fwdbwd = flops_per_token * T\n",
        "        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter\n",
        "        # express our flops throughput as ratio of A100 bfloat16 peak flops\n",
        "        flops_achieved = flops_per_iter * (1.0 / dt)  # per second\n",
        "        flops_promised = 312e12  # A100 GPU bfloat16 peak flops is 312 TFLOPS\n",
        "        mfu = flops_achieved / flops_promised\n",
        "        return mfu\n",
        "\n",
        "    def forwards(self, idx, targets=None):\n",
        "        device = idx.device\n",
        "        b, t = idx.size()\n",
        "        assert (\n",
        "            t <= self.config.block_size\n",
        "        ), f\"Cannot forward sequence of length {t}, block size is only {self.config.block_size}\"\n",
        "        pos = torch.empty((1, t), dtype=torch.long, device=device)\n",
        "\n",
        "        for i in range(t):\n",
        "            pos[0, i] = i\n",
        "\n",
        "        # # # forward the GPT model itself\n",
        "        # token embeddings of shape (b, t, n_embd), idx -> token_emb\n",
        "        idx = self.transformer.wte(idx)\n",
        "        # position embeddings of shape (1, t, n_embd)\n",
        "        pos_emb = self.transformer.wpe(pos)\n",
        "\n",
        "        idx = self.transformer.drop(idx + pos_emb)\n",
        "\n",
        "        for block in self.transformer.h:\n",
        "            idx = block(idx)\n",
        "\n",
        "        idx = self.transformer.ln_f(idx)\n",
        "\n",
        "        if targets is not None:\n",
        "            # if we are given some desired targets also calculate the loss\n",
        "            idx = self.lm_head(idx)\n",
        "            loss = F.cross_entropy(\n",
        "                idx.view(-1, idx.size(-1)), targets.view(-1), ignore_index=-1\n",
        "            )\n",
        "        else:\n",
        "            # inference-time mini-optimization: only forward the lm_head on the very last position\n",
        "            idx = self.lm_head(\n",
        "                idx[:, [-1], :]\n",
        "            )  # note: using list [-1] to preserve the time dim\n",
        "            loss = None\n",
        "\n",
        "        return idx, loss\n",
        "\n",
        "    def forward(self, idx, targets=None):\n",
        "        \"\"\"\n",
        "        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete\n",
        "        the sequence max_new_tokens times, feeding the predictions back into the model each time.\n",
        "        Most likely you'll want to make sure to be in model.eval() mode of operation for this.\n",
        "        \"\"\"\n",
        "        if self.train_mode:\n",
        "            return self.forwards(\n",
        "                idx, targets\n",
        "            )  # When training don't override the forward method\n",
        "        original_size = idx.size(1)\n",
        "        for _ in range(5):\n",
        "        # if the sequence context is growing too long we must crop it at block_size\n",
        "            idx_cond = (\n",
        "                idx\n",
        "                if idx.size(1) <= self.config.block_size\n",
        "                else idx[:, -self.config.block_size :]\n",
        "            )\n",
        "            logits, _ = self.forwards(idx_cond)\n",
        "            # pluck the logits at the final step\n",
        "            logits = logits[:, -1, :]\n",
        "\n",
        "            idx = torch.cat((idx, torch.argmax(logits, dim=1, keepdim=True)), dim=1)\n",
        "        return idx[:, original_size:]\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "qYc38uTvpktW"
      },
      "source": [
        "### Prepare Dataset"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 3,
      "metadata": {
        "id": "C-SNhvuVJtyr",
        "outputId": "f97e6bfa-0c4b-4b3a-883e-42f5f4210451",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "length of dataset in characters: 1,115,394\n",
            "all the unique characters: \n",
            " !$&',-.3:;?ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz\n",
            "vocab size: 65\n",
            "train has 1,003,854 tokens\n",
            "val has 111,540 tokens\n"
          ]
        }
      ],
      "source": [
        "\"\"\"\n",
        "Prepare the Shakespeare dataset for character-level language modeling.\n",
        "So instead of encoding with GPT-2 BPE tokens, we just map characters to ints.\n",
        "Will save train.bin, val.bin containing the ids, and meta.pkl containing the\n",
        "encoder and decoder and some other related info.\n",
        "\"\"\"\n",
        "import os\n",
        "import pickle\n",
        "import requests\n",
        "import numpy as np\n",
        "\n",
        "# download the tiny shakespeare dataset\n",
        "\n",
        "DATA_DIR=\"./data\"\n",
        "FILE_PATH=os.path.join(DATA_DIR, 'shakespeare.txt')\n",
        "if not os.path.exists(DATA_DIR):\n",
        "  os.mkdir(DATA_DIR)\n",
        "data_url = 'https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt'\n",
        "with open(FILE_PATH, 'w') as f:\n",
        "    f.write(requests.get(data_url).text)\n",
        "\n",
        "with open(FILE_PATH, 'r') as f:\n",
        "    data = f.read()\n",
        "print(f\"length of dataset in characters: {len(data):,}\")\n",
        "\n",
        "# get all the unique characters that occur in this text\n",
        "chars = sorted(list(set(data)))\n",
        "vocab_size = len(chars)\n",
        "print(\"all the unique characters:\", ''.join(chars))\n",
        "print(f\"vocab size: {vocab_size:,}\")\n",
        "\n",
        "# create a mapping from characters to integers\n",
        "stoi = { ch:i for i,ch in enumerate(chars) }\n",
        "itos = { i:ch for i,ch in enumerate(chars) }\n",
        "def encode(s):\n",
        "    return [stoi[c] for c in s] # encoder: take a string, output a list of integers\n",
        "def decode(l):\n",
        "    return ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string\n",
        "\n",
        "# create the train and test splits\n",
        "n = len(data)\n",
        "train_data = data[:int(n*0.9)]\n",
        "val_data = data[int(n*0.9):]\n",
        "\n",
        "# encode both to integers\n",
        "train_ids = encode(train_data)\n",
        "val_ids = encode(val_data)\n",
        "print(f\"train has {len(train_ids):,} tokens\")\n",
        "print(f\"val has {len(val_ids):,} tokens\")\n",
        "\n",
        "# export to bin files\n",
        "train_ids = np.array(train_ids, dtype=np.uint16)\n",
        "val_ids = np.array(val_ids, dtype=np.uint16)\n",
        "train_ids.tofile(os.path.join(DATA_DIR, 'train.bin'))\n",
        "val_ids.tofile(os.path.join(DATA_DIR, 'val.bin'))\n",
        "\n",
        "# save the meta information as well, to help us encode/decode later\n",
        "meta = {\n",
        "    'vocab_size': vocab_size,\n",
        "    'itos': itos,\n",
        "    'stoi': stoi,\n",
        "}\n",
        "with open(os.path.join(DATA_DIR, 'meta.pkl'), 'wb') as f:\n",
        "    pickle.dump(meta, f)\n",
        "\n",
        "# length of dataset in characters:  1115394\n",
        "# all the unique characters:\n",
        "#  !$&',-.3:;?ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz\n",
        "# vocab size: 65\n",
        "# train has 1003854 tokens\n",
        "# val has 111540 tokens\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "eoecPp3jpktW"
      },
      "source": [
        "### Train the model"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {
        "id": "EN1ItHTBLxSi",
        "outputId": "7e952778-da88-4eda-8285-a00c990dfa9e",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "tokens per iteration will be: 1,024\n",
            "found vocab_size = 65 (inside data/meta.pkl)\n",
            "Initializing a new model from scratch\n",
            "number of parameters: 0.34M\n",
            "num decayed parameter tensors: 14, with 341,088 parameters\n",
            "num non-decayed parameter tensors: 7, with 672 parameters\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "/usr/local/lib/python3.10/dist-packages/torch/cuda/amp/grad_scaler.py:125: UserWarning: torch.cuda.amp.GradScaler is enabled, but CUDA is not available.  Disabling.\n",
            "  warnings.warn(\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "using fused AdamW: False\n",
            "compiling the model... (takes a ~minute)\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "[2024-02-01 20:42:19,295] [0/0] torch._inductor.fx_passes.split_cat: [WARNING] example value absent for node: split_1\n",
            "[2024-02-01 20:42:19,296] [0/0] torch._inductor.fx_passes.split_cat: [WARNING] example value absent for node: split\n",
            "No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "step 0: train loss 4.1708, val loss 4.1664\n",
            "iter 0: loss 4.1697, time 21488.29ms, mfu -100.00%\n",
            "iter 1: loss 4.1753, time 17.19ms, mfu -100.00%\n",
            "iter 2: loss 4.1557, time 15.86ms, mfu -100.00%\n",
            "iter 3: loss 4.1560, time 15.54ms, mfu -100.00%\n",
            "iter 4: loss 4.1479, time 15.21ms, mfu -100.00%\n",
            "iter 5: loss 4.1254, time 15.30ms, mfu 0.05%\n",
            "iter 6: loss 4.0951, time 15.15ms, mfu 0.05%\n",
            "iter 7: loss 4.0726, time 15.32ms, mfu 0.05%\n",
            "iter 8: loss 4.0357, time 14.90ms, mfu 0.05%\n",
            "iter 9: loss 4.0177, time 14.91ms, mfu 0.05%\n",
            "iter 10: loss 3.9766, time 15.21ms, mfu 0.05%\n",
            "iter 11: loss 3.9375, time 15.17ms, mfu 0.05%\n",
            "iter 12: loss 3.9252, time 15.75ms, mfu 0.05%\n",
            "iter 13: loss 3.8665, time 17.39ms, mfu 0.04%\n",
            "iter 14: loss 3.8533, time 15.29ms, mfu 0.04%\n",
            "iter 15: loss 3.8250, time 15.39ms, mfu 0.04%\n",
            "iter 16: loss 3.8300, time 15.14ms, mfu 0.04%\n",
            "iter 17: loss 3.7883, time 14.83ms, mfu 0.05%\n",
            "iter 18: loss 3.7875, time 14.84ms, mfu 0.05%\n",
            "iter 19: loss 3.7742, time 15.12ms, mfu 0.05%\n",
            "iter 20: loss 3.7545, time 14.79ms, mfu 0.05%\n",
            "iter 21: loss 3.7345, time 15.46ms, mfu 0.05%\n",
            "iter 22: loss 3.7235, time 15.04ms, mfu 0.05%\n",
            "iter 23: loss 3.7051, time 15.17ms, mfu 0.05%\n",
            "iter 24: loss 3.7472, time 15.08ms, mfu 0.05%\n",
            "iter 25: loss 3.6630, time 15.08ms, mfu 0.05%\n",
            "iter 26: loss 3.6455, time 15.11ms, mfu 0.05%\n",
            "iter 27: loss 3.6514, time 15.51ms, mfu 0.05%\n",
            "iter 28: loss 3.6051, time 15.32ms, mfu 0.05%\n",
            "iter 29: loss 3.5626, time 16.80ms, mfu 0.04%\n",
            "iter 30: loss 3.5923, time 15.44ms, mfu 0.04%\n",
            "iter 31: loss 3.5824, time 15.12ms, mfu 0.05%\n",
            "iter 32: loss 3.5221, time 15.37ms, mfu 0.05%\n",
            "iter 33: loss 3.5523, time 14.84ms, mfu 0.05%\n",
            "iter 34: loss 3.5210, time 14.72ms, mfu 0.05%\n",
            "iter 35: loss 3.4956, time 15.43ms, mfu 0.05%\n",
            "iter 36: loss 3.4522, time 15.91ms, mfu 0.05%\n",
            "iter 37: loss 3.4721, time 15.46ms, mfu 0.05%\n",
            "iter 38: loss 3.4638, time 15.10ms, mfu 0.05%\n",
            "iter 39: loss 3.4700, time 15.40ms, mfu 0.05%\n",
            "iter 40: loss 3.3838, time 15.24ms, mfu 0.05%\n",
            "iter 41: loss 3.3406, time 15.55ms, mfu 0.05%\n",
            "iter 42: loss 3.3452, time 15.31ms, mfu 0.05%\n",
            "iter 43: loss 3.2886, time 15.00ms, mfu 0.05%\n",
            "iter 44: loss 3.2659, time 15.06ms, mfu 0.05%\n",
            "iter 45: loss 3.2697, time 15.26ms, mfu 0.05%\n",
            "iter 46: loss 3.3064, time 15.34ms, mfu 0.05%\n",
            "iter 47: loss 3.2620, time 15.12ms, mfu 0.05%\n",
            "iter 48: loss 3.2432, time 15.08ms, mfu 0.05%\n",
            "iter 49: loss 3.1986, time 15.14ms, mfu 0.05%\n",
            "iter 50: loss 3.1692, time 15.37ms, mfu 0.05%\n",
            "iter 51: loss 3.2101, time 15.16ms, mfu 0.05%\n",
            "iter 52: loss 3.1993, time 15.45ms, mfu 0.05%\n",
            "iter 53: loss 3.1200, time 15.18ms, mfu 0.05%\n",
            "iter 54: loss 3.1264, time 14.88ms, mfu 0.05%\n",
            "iter 55: loss 3.0649, time 15.56ms, mfu 0.05%\n",
            "iter 56: loss 3.0429, time 15.03ms, mfu 0.05%\n",
            "iter 57: loss 3.0382, time 14.65ms, mfu 0.05%\n",
            "iter 58: loss 3.0583, time 15.52ms, mfu 0.05%\n",
            "iter 59: loss 2.9773, time 14.92ms, mfu 0.05%\n",
            "iter 60: loss 3.0380, time 14.77ms, mfu 0.05%\n",
            "iter 61: loss 2.9201, time 14.70ms, mfu 0.05%\n",
            "iter 62: loss 3.0220, time 15.37ms, mfu 0.05%\n",
            "iter 63: loss 2.9231, time 14.92ms, mfu 0.05%\n",
            "iter 64: loss 2.9820, time 15.20ms, mfu 0.05%\n",
            "iter 65: loss 2.9536, time 15.42ms, mfu 0.05%\n",
            "iter 66: loss 2.9140, time 15.52ms, mfu 0.05%\n",
            "iter 67: loss 2.9347, time 14.75ms, mfu 0.05%\n",
            "iter 68: loss 2.9079, time 15.15ms, mfu 0.05%\n",
            "iter 69: loss 2.9735, time 15.69ms, mfu 0.05%\n",
            "iter 70: loss 2.8604, time 15.39ms, mfu 0.05%\n",
            "iter 71: loss 2.8745, time 15.27ms, mfu 0.05%\n",
            "iter 72: loss 2.8085, time 16.62ms, mfu 0.05%\n",
            "iter 73: loss 2.8854, time 15.58ms, mfu 0.04%\n",
            "iter 74: loss 2.8348, time 15.17ms, mfu 0.05%\n",
            "iter 75: loss 2.8814, time 15.23ms, mfu 0.05%\n",
            "iter 76: loss 2.8533, time 18.14ms, mfu 0.04%\n",
            "iter 77: loss 2.8558, time 18.04ms, mfu 0.04%\n",
            "iter 78: loss 2.8652, time 18.71ms, mfu 0.04%\n",
            "iter 79: loss 2.8195, time 17.59ms, mfu 0.04%\n",
            "iter 80: loss 2.7054, time 18.55ms, mfu 0.04%\n",
            "iter 81: loss 2.7841, time 19.25ms, mfu 0.04%\n",
            "iter 82: loss 2.9033, time 18.42ms, mfu 0.04%\n",
            "iter 83: loss 2.8279, time 18.56ms, mfu 0.04%\n",
            "iter 84: loss 2.7585, time 19.22ms, mfu 0.04%\n",
            "iter 85: loss 2.6999, time 18.99ms, mfu 0.04%\n",
            "iter 86: loss 2.7767, time 18.93ms, mfu 0.04%\n",
            "iter 87: loss 2.7366, time 18.21ms, mfu 0.04%\n",
            "iter 88: loss 2.7877, time 18.51ms, mfu 0.04%\n",
            "iter 89: loss 2.7298, time 15.70ms, mfu 0.04%\n",
            "iter 90: loss 2.7377, time 14.95ms, mfu 0.04%\n",
            "iter 91: loss 2.6334, time 15.00ms, mfu 0.04%\n",
            "iter 92: loss 2.6666, time 15.63ms, mfu 0.04%\n",
            "iter 93: loss 2.6042, time 15.58ms, mfu 0.04%\n",
            "iter 94: loss 2.7643, time 15.07ms, mfu 0.04%\n",
            "iter 95: loss 2.7243, time 15.10ms, mfu 0.04%\n",
            "iter 96: loss 2.7619, time 15.05ms, mfu 0.04%\n",
            "iter 97: loss 2.6290, time 15.09ms, mfu 0.04%\n",
            "iter 98: loss 2.6788, time 15.12ms, mfu 0.04%\n",
            "iter 99: loss 2.7000, time 15.28ms, mfu 0.04%\n",
            "iter 100: loss 2.6518, time 14.99ms, mfu 0.04%\n",
            "iter 101: loss 2.7560, time 15.40ms, mfu 0.04%\n",
            "iter 102: loss 2.6619, time 15.28ms, mfu 0.04%\n",
            "iter 103: loss 2.6390, time 15.98ms, mfu 0.04%\n",
            "iter 104: loss 2.6788, time 15.24ms, mfu 0.04%\n",
            "iter 105: loss 2.6600, time 14.89ms, mfu 0.04%\n",
            "iter 106: loss 2.7043, time 15.02ms, mfu 0.04%\n",
            "iter 107: loss 2.6232, time 15.91ms, mfu 0.04%\n",
            "iter 108: loss 2.6425, time 15.49ms, mfu 0.04%\n",
            "iter 109: loss 2.6943, time 15.43ms, mfu 0.04%\n",
            "iter 110: loss 2.6045, time 16.15ms, mfu 0.04%\n",
            "iter 111: loss 2.6670, time 15.05ms, mfu 0.04%\n",
            "iter 112: loss 2.5576, time 15.31ms, mfu 0.04%\n",
            "iter 113: loss 2.6802, time 15.41ms, mfu 0.04%\n",
            "iter 114: loss 2.6305, time 14.98ms, mfu 0.04%\n",
            "iter 115: loss 2.6010, time 15.33ms, mfu 0.04%\n",
            "iter 116: loss 2.5796, time 15.59ms, mfu 0.04%\n",
            "iter 117: loss 2.5308, time 15.64ms, mfu 0.04%\n",
            "iter 118: loss 2.5855, time 14.59ms, mfu 0.04%\n",
            "iter 119: loss 2.6275, time 14.87ms, mfu 0.05%\n",
            "iter 120: loss 2.6243, time 14.87ms, mfu 0.05%\n",
            "iter 121: loss 2.6038, time 15.92ms, mfu 0.05%\n",
            "iter 122: loss 2.6008, time 16.32ms, mfu 0.04%\n",
            "iter 123: loss 2.5022, time 19.35ms, mfu 0.04%\n",
            "iter 124: loss 2.5295, time 18.92ms, mfu 0.04%\n",
            "iter 125: loss 2.5948, time 18.99ms, mfu 0.04%\n",
            "iter 126: loss 2.5400, time 18.75ms, mfu 0.04%\n",
            "iter 127: loss 2.6712, time 21.15ms, mfu 0.04%\n",
            "iter 128: loss 2.5978, time 19.33ms, mfu 0.04%\n",
            "iter 129: loss 2.5710, time 18.93ms, mfu 0.04%\n",
            "iter 130: loss 2.5817, time 20.67ms, mfu 0.04%\n",
            "iter 131: loss 2.5699, time 19.40ms, mfu 0.04%\n",
            "iter 132: loss 2.4486, time 16.75ms, mfu 0.04%\n",
            "iter 133: loss 2.5532, time 15.96ms, mfu 0.04%\n",
            "iter 134: loss 2.5778, time 15.20ms, mfu 0.04%\n",
            "iter 135: loss 2.5059, time 15.24ms, mfu 0.04%\n",
            "iter 136: loss 2.5254, time 15.12ms, mfu 0.04%\n",
            "iter 137: loss 2.4987, time 15.82ms, mfu 0.04%\n",
            "iter 138: loss 2.5988, time 19.09ms, mfu 0.04%\n",
            "iter 139: loss 2.5627, time 17.55ms, mfu 0.04%\n",
            "iter 140: loss 2.5481, time 18.09ms, mfu 0.04%\n",
            "iter 141: loss 2.5456, time 18.29ms, mfu 0.04%\n",
            "iter 142: loss 2.5845, time 18.70ms, mfu 0.04%\n",
            "iter 143: loss 2.5867, time 18.02ms, mfu 0.04%\n",
            "iter 144: loss 2.5249, time 17.67ms, mfu 0.04%\n",
            "iter 145: loss 2.5530, time 17.23ms, mfu 0.04%\n",
            "iter 146: loss 2.4775, time 17.61ms, mfu 0.04%\n",
            "iter 147: loss 2.4726, time 18.08ms, mfu 0.04%\n",
            "iter 148: loss 2.4967, time 18.98ms, mfu 0.04%\n",
            "iter 149: loss 2.5593, time 17.28ms, mfu 0.04%\n",
            "iter 150: loss 2.5603, time 17.69ms, mfu 0.04%\n",
            "iter 151: loss 2.5385, time 18.51ms, mfu 0.04%\n",
            "iter 152: loss 2.5594, time 17.85ms, mfu 0.04%\n",
            "iter 153: loss 2.5166, time 17.81ms, mfu 0.04%\n",
            "iter 154: loss 2.4737, time 19.76ms, mfu 0.04%\n",
            "iter 155: loss 2.5284, time 18.64ms, mfu 0.04%\n",
            "iter 156: loss 2.4409, time 18.76ms, mfu 0.04%\n",
            "iter 157: loss 2.4837, time 17.60ms, mfu 0.04%\n",
            "iter 158: loss 2.5167, time 18.00ms, mfu 0.04%\n",
            "iter 159: loss 2.5171, time 18.28ms, mfu 0.04%\n",
            "iter 160: loss 2.5408, time 17.96ms, mfu 0.04%\n",
            "iter 161: loss 2.5084, time 19.12ms, mfu 0.04%\n",
            "iter 162: loss 2.4151, time 19.14ms, mfu 0.04%\n",
            "iter 163: loss 2.5133, time 16.03ms, mfu 0.04%\n",
            "iter 164: loss 2.4398, time 16.15ms, mfu 0.04%\n",
            "iter 165: loss 2.4783, time 16.02ms, mfu 0.04%\n",
            "iter 166: loss 2.5284, time 15.82ms, mfu 0.04%\n",
            "iter 167: loss 2.4459, time 15.65ms, mfu 0.04%\n",
            "iter 168: loss 2.4881, time 16.31ms, mfu 0.04%\n",
            "iter 169: loss 2.4807, time 15.60ms, mfu 0.04%\n",
            "iter 170: loss 2.4425, time 15.58ms, mfu 0.04%\n",
            "iter 171: loss 2.4770, time 15.63ms, mfu 0.04%\n",
            "iter 172: loss 2.4737, time 15.31ms, mfu 0.04%\n",
            "iter 173: loss 2.4802, time 15.97ms, mfu 0.04%\n",
            "iter 174: loss 2.5004, time 15.44ms, mfu 0.04%\n",
            "iter 175: loss 2.5208, time 15.47ms, mfu 0.04%\n",
            "iter 176: loss 2.4841, time 15.64ms, mfu 0.04%\n",
            "iter 177: loss 2.4494, time 15.28ms, mfu 0.04%\n",
            "iter 178: loss 2.5139, time 15.44ms, mfu 0.04%\n",
            "iter 179: loss 2.4887, time 15.08ms, mfu 0.04%\n",
            "iter 180: loss 2.4871, time 16.31ms, mfu 0.04%\n",
            "iter 181: loss 2.4706, time 16.25ms, mfu 0.04%\n",
            "iter 182: loss 2.5013, time 15.31ms, mfu 0.04%\n",
            "iter 183: loss 2.5103, time 14.79ms, mfu 0.04%\n",
            "iter 184: loss 2.4708, time 14.97ms, mfu 0.04%\n",
            "iter 185: loss 2.4787, time 14.86ms, mfu 0.04%\n",
            "iter 186: loss 2.4228, time 15.06ms, mfu 0.04%\n",
            "iter 187: loss 2.4344, time 15.14ms, mfu 0.04%\n",
            "iter 188: loss 2.4265, time 15.14ms, mfu 0.04%\n",
            "iter 189: loss 2.4444, time 15.44ms, mfu 0.04%\n",
            "iter 190: loss 2.4519, time 15.20ms, mfu 0.04%\n",
            "iter 191: loss 2.5300, time 15.25ms, mfu 0.04%\n",
            "iter 192: loss 2.4184, time 15.22ms, mfu 0.04%\n",
            "iter 193: loss 2.4493, time 15.33ms, mfu 0.04%\n",
            "iter 194: loss 2.4406, time 15.72ms, mfu 0.04%\n",
            "iter 195: loss 2.5127, time 16.07ms, mfu 0.04%\n",
            "iter 196: loss 2.4863, time 15.31ms, mfu 0.04%\n",
            "iter 197: loss 2.4092, time 15.08ms, mfu 0.04%\n",
            "iter 198: loss 2.5352, time 14.94ms, mfu 0.04%\n",
            "iter 199: loss 2.5001, time 16.64ms, mfu 0.04%\n",
            "step 200: train loss 2.4671, val loss 2.4676\n",
            "saving checkpoint to pretrained\n",
            "iter 200: loss 2.4297, time 1919.96ms, mfu 0.04%\n"
          ]
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 1000x500 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ],
      "source": [
        "\n",
        "\"\"\"\n",
        "This training script can be run both on a single gpu in debug mode,\n",
        "and also in a larger training run with distributed data parallel (ddp).\n",
        "\n",
        "To run on a single GPU, example:\n",
        "$ python train.py --batch_size=32 --compile=False\n",
        "\n",
        "To run with DDP on 4 gpus on 1 node, example:\n",
        "$ torchrun --standalone --nproc_per_node=4 train.py\n",
        "\n",
        "To run with DDP on 4 gpus across 2 nodes, example:\n",
        "- Run on the first (master) node with example IP 123.456.123.456:\n",
        "$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr=123.456.123.456 --master_port=1234 train.py\n",
        "- Run on the worker node:\n",
        "$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr=123.456.123.456 --master_port=1234 train.py\n",
        "(If your cluster does not have Infiniband interconnect prepend NCCL_IB_DISABLE=1)\n",
        "\"\"\"\n",
        "\n",
        "import os\n",
        "import time\n",
        "import math\n",
        "import pickle\n",
        "from contextlib import nullcontext\n",
        "from matplotlib import pyplot as plt\n",
        "from threading import Thread\n",
        "\n",
        "import numpy as np\n",
        "import torch\n",
        "from torch.nn.parallel import DistributedDataParallel as DDP\n",
        "from torch.distributed import init_process_group, destroy_process_group\n",
        "\n",
        "train_losses = []\n",
        "val_losses = []\n",
        "\n",
        "# -----------------------------------------------------------------------------\n",
        "# default config values designed to train a gpt2 (124M) on OpenWebText\n",
        "# I/O\n",
        "out_dir = 'pretrained'\n",
        "eval_interval = 200\n",
        "log_interval = 1\n",
        "eval_iters = 20\n",
        "eval_only = False # if True, script exits right after the first eval\n",
        "always_save_checkpoint = True # if True, always save a checkpoint after each eval\n",
        "init_from = 'scratch' # 'scratch' or 'resume' or 'gpt2*'\n",
        "# wandb logging\n",
        "wandb_log = False # disabled by default\n",
        "wandb_project = 'owt'\n",
        "wandb_run_name = 'gpt2' # 'run' + str(time.time())\n",
        "# data\n",
        "dataset = 'openwebtext'\n",
        "gradient_accumulation_steps = 5 * 8 # used to simulate larger batch sizes\n",
        "batch_size = 12 # if gradient_accumulation_steps > 1, this is the micro-batch size\n",
        "block_size = 64\n",
        "# model\n",
        "n_layer = 4\n",
        "n_head = 4\n",
        "n_embd = 128\n",
        "dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+\n",
        "bias = False # do we use bias inside LayerNorm and Linear layers?\n",
        "# adamw optimizer\n",
        "learning_rate = 6e-4 # max learning rate\n",
        "max_iters = 600000 # total number of training iterations\n",
        "weight_decay = 1e-1\n",
        "beta1 = 0.9\n",
        "beta2 = 0.95\n",
        "grad_clip = 0.0 # clip gradients at this value, or disable if == 0.0\n",
        "# learning rate decay settings\n",
        "decay_lr = True # whether to decay the learning rate\n",
        "warmup_iters = 100 # how many steps to warm up for\n",
        "lr_decay_iters = 600000 # should be ~= max_iters per Chinchilla\n",
        "min_lr = 6e-5 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla\n",
        "# DDP settings\n",
        "backend = 'nccl' # 'nccl', 'gloo', etc.\n",
        "# system\n",
        "device = 'cpu' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks\n",
        "dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler\n",
        "compile = True # use PyTorch 2.0 to compile the model to be faster\n",
        "# -----------------------------------------------------------------------------\n",
        "config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]\n",
        "config = {k: globals()[k] for k in config_keys} # will be useful for logging\n",
        "# -----------------------------------------------------------------------------\n",
        "# train a miniature character-level shakespeare model\n",
        "# good for debugging and playing on macbooks and such\n",
        "\n",
        "out_dir = 'pretrained'\n",
        "eval_interval = 200 # keep frequent because we'll overfit\n",
        "eval_iters = 200\n",
        "log_interval = 1 # don't print too too often\n",
        "\n",
        "# we expect to overfit on this small dataset, so only save when val improves\n",
        "always_save_checkpoint = True\n",
        "\n",
        "wandb_log = False # override via command line if you like\n",
        "wandb_project = 'shakespeare-char'\n",
        "wandb_run_name = 'mini-gpt'\n",
        "\n",
        "dataset = 'data'\n",
        "gradient_accumulation_steps = 1\n",
        "batch_size = 32\n",
        "block_size = 32 # context of up to 256 previous characters\n",
        "\n",
        "# baby GPT model :)\n",
        "n_layer = 2\n",
        "n_head = 2\n",
        "n_embd = 96\n",
        "dropout = 0.0\n",
        "\n",
        "learning_rate = 1e-3 # with baby networks can afford to go a bit higher\n",
        "max_iters = 200\n",
        "lr_decay_iters = 200 # make equal to max_iters usually\n",
        "min_lr = 1e-4 # learning_rate / 10 usually\n",
        "beta2 = 0.99 # make a bit bigger because number of tokens per iter is small\n",
        "\n",
        "warmup_iters = 100 # not super necessary potentially\n",
        "\n",
        "# on macbook also add\n",
        "# device = 'mps'  # run on cpu only\n",
        "# compile = False # do not torch compile the model\n",
        "\n",
        "\n",
        "# various inits, derived attributes, I/O setup\n",
        "ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?\n",
        "if ddp:\n",
        "    init_process_group(backend=backend)\n",
        "    ddp_rank = int(os.environ['RANK'])\n",
        "    ddp_local_rank = int(os.environ['LOCAL_RANK'])\n",
        "    ddp_world_size = int(os.environ['WORLD_SIZE'])\n",
        "    device = f'cuda:{ddp_local_rank}'\n",
        "    torch.cuda.set_device(device)\n",
        "    master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.\n",
        "    seed_offset = ddp_rank # each process gets a different seed\n",
        "    # world_size number of processes will be training simultaneously, so we can scale\n",
        "    # down the desired gradient accumulation iterations per process proportionally\n",
        "    assert gradient_accumulation_steps % ddp_world_size == 0\n",
        "    gradient_accumulation_steps //= ddp_world_size\n",
        "else:\n",
        "    # if not ddp, we are running on a single gpu, and one process\n",
        "    master_process = True\n",
        "    seed_offset = 0\n",
        "    ddp_world_size = 1\n",
        "tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size\n",
        "print(f\"tokens per iteration will be: {tokens_per_iter:,}\")\n",
        "\n",
        "if master_process:\n",
        "    os.makedirs(out_dir, exist_ok=True)\n",
        "torch.manual_seed(1337 + seed_offset)\n",
        "torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul\n",
        "torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn\n",
        "device_type = 'cuda' if 'cuda' in device else 'mps' # for later use in torch.autocast\n",
        "# note: float16 data type will automatically use a GradScaler\n",
        "ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]\n",
        "ctx = nullcontext()\n",
        "\n",
        "# poor man's data loader\n",
        "data_dir = os.path.join(dataset)\n",
        "train_data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')\n",
        "val_data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')\n",
        "def get_batch(split):\n",
        "    data = train_data if split == 'train' else val_data\n",
        "    ix = torch.randint(len(data) - block_size, (batch_size,))\n",
        "    x = torch.stack([torch.from_numpy((data[i:i+block_size]).astype(np.int64)) for i in ix])\n",
        "    y = torch.stack([torch.from_numpy((data[i+1:i+1+block_size]).astype(np.int64)) for i in ix])\n",
        "    if device_type == 'cuda':\n",
        "        # pin arrays x,y, which allows us to move them to GPU asynchronously (non_blocking=True)\n",
        "        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)\n",
        "    else:\n",
        "        x, y = x.to(device), y.to(device)\n",
        "    return x, y\n",
        "\n",
        "# init these up here, can override if init_from='resume' (i.e. from a checkpoint)\n",
        "iter_num = 0\n",
        "best_val_loss = 1e9\n",
        "\n",
        "# attempt to derive vocab_size from the dataset\n",
        "meta_path = os.path.join(data_dir, 'meta.pkl')\n",
        "meta_vocab_size = None\n",
        "if os.path.exists(meta_path):\n",
        "    with open(meta_path, 'rb') as f:\n",
        "        meta = pickle.load(f)\n",
        "    meta_vocab_size = meta['vocab_size']\n",
        "    print(f\"found vocab_size = {meta_vocab_size} (inside {meta_path})\")\n",
        "\n",
        "# model init\n",
        "model_args = dict(n_layer=n_layer, n_head=n_head, n_embd=n_embd, block_size=block_size,\n",
        "                  bias=bias, vocab_size=None, dropout=dropout) # start with model_args from command line\n",
        "if init_from == 'scratch':\n",
        "    # init a new model from scratch\n",
        "    print(\"Initializing a new model from scratch\")\n",
        "    # determine the vocab size we'll use for from-scratch training\n",
        "    if meta_vocab_size is None:\n",
        "        print(\"defaulting to vocab_size of GPT-2 to 50304 (50257 rounded up for efficiency)\")\n",
        "    model_args['vocab_size'] = meta_vocab_size if meta_vocab_size is not None else 50304\n",
        "    gptconf = GPTConfig(**model_args)\n",
        "    model = GPT(gptconf, True)\n",
        "elif init_from == 'resume':\n",
        "    print(f\"Resuming training from {out_dir}\")\n",
        "    # resume training from a checkpoint.\n",
        "    ckpt_path = os.path.join(out_dir, 'ckpt.pt')\n",
        "    checkpoint = torch.load(ckpt_path, map_location=device)\n",
        "    checkpoint_model_args = checkpoint['model_args']\n",
        "    # force these config attributes to be equal otherwise we can't even resume training\n",
        "    # the rest of the attributes (e.g. dropout) can stay as desired from command line\n",
        "    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:\n",
        "        model_args[k] = checkpoint_model_args[k]\n",
        "    # create the model\n",
        "    gptconf = GPTConfig(**model_args)\n",
        "    model = GPT(gptconf)\n",
        "    state_dict = checkpoint['model']\n",
        "    # fix the keys of the state dictionary :(\n",
        "    # honestly no idea how checkpoints sometimes get this prefix, have to debug more\n",
        "    unwanted_prefix = '_orig_mod.'\n",
        "    for k,v in list(state_dict.items()):\n",
        "        if k.startswith(unwanted_prefix):\n",
        "            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)\n",
        "    model.load_state_dict(state_dict)\n",
        "    iter_num = checkpoint['iter_num']\n",
        "    best_val_loss = checkpoint['best_val_loss']\n",
        "elif init_from.startswith('gpt2'):\n",
        "    print(f\"Initializing from OpenAI GPT-2 weights: {init_from}\")\n",
        "    # initialize from OpenAI GPT-2 weights\n",
        "    override_args = dict(dropout=dropout)\n",
        "    model = GPT.from_pretrained(init_from, override_args)\n",
        "    # read off the created config params, so we can store them into checkpoint correctly\n",
        "    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:\n",
        "        model_args[k] = getattr(model.config, k)\n",
        "# crop down the model block size if desired, using model surgery\n",
        "if block_size < model.config.block_size:\n",
        "    model.crop_block_size(block_size)\n",
        "    model_args['block_size'] = block_size # so that the checkpoint will have the right value\n",
        "model.to(device)\n",
        "\n",
        "# initialize a GradScaler. If enabled=False scaler is a no-op\n",
        "scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))\n",
        "\n",
        "# optimizer\n",
        "optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)\n",
        "if init_from == 'resume':\n",
        "    optimizer.load_state_dict(checkpoint['optimizer'])\n",
        "checkpoint = None # free up memory\n",
        "\n",
        "# compile the model\n",
        "if compile:\n",
        "    print(\"compiling the model... (takes a ~minute)\")\n",
        "    unoptimized_model = model\n",
        "    model = torch.compile(model) # requires PyTorch 2.0\n",
        "\n",
        "# wrap model into DDP container\n",
        "if ddp:\n",
        "    model = DDP(model, device_ids=[ddp_local_rank])\n",
        "\n",
        "# helps estimate an arbitrarily accurate loss over either split using many batches\n",
        "@torch.no_grad()\n",
        "def estimate_loss():\n",
        "    out = {}\n",
        "    model.eval()\n",
        "    for split in ['train', 'val']:\n",
        "        losses = torch.zeros(eval_iters)\n",
        "        for k in range(eval_iters):\n",
        "            X, Y = get_batch(split)\n",
        "            with ctx:\n",
        "                logits, loss = model(X, Y)\n",
        "            losses[k] = loss.item()\n",
        "        out[split] = losses.mean()\n",
        "    train_losses.append(out['train'])\n",
        "    val_losses.append(out['val'])\n",
        "    model.train()\n",
        "    return out\n",
        "\n",
        "# learning rate decay scheduler (cosine with warmup)\n",
        "def get_lr(it):\n",
        "    # 1) linear warmup for warmup_iters steps\n",
        "    if it < warmup_iters:\n",
        "        return learning_rate * it / warmup_iters\n",
        "    # 2) if it > lr_decay_iters, return min learning rate\n",
        "    if it > lr_decay_iters:\n",
        "        return min_lr\n",
        "    # 3) in between, use cosine decay down to min learning rate\n",
        "    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)\n",
        "    assert 0 <= decay_ratio <= 1\n",
        "    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1\n",
        "    return min_lr + coeff * (learning_rate - min_lr)\n",
        "\n",
        "# logging\n",
        "if wandb_log and master_process:\n",
        "    import wandb\n",
        "    wandb.init(project=wandb_project, name=wandb_run_name, config=config)\n",
        "\n",
        "# training loop\n",
        "X, Y = get_batch('train') # fetch the very first batch\n",
        "t0 = time.time()\n",
        "local_iter_num = 0 # number of iterations in the lifetime of this process\n",
        "raw_model = model.module if ddp else model # unwrap DDP container if needed\n",
        "running_mfu = -1.0\n",
        "while True:\n",
        "\n",
        "    # determine and set the learning rate for this iteration\n",
        "    lr = get_lr(iter_num) if decay_lr else learning_rate\n",
        "    for param_group in optimizer.param_groups:\n",
        "        param_group['lr'] = lr\n",
        "\n",
        "    # evaluate the loss on train/val sets and write checkpoints\n",
        "    if iter_num % eval_interval == 0 and master_process:\n",
        "        losses = estimate_loss()\n",
        "\n",
        "        print(f\"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}\")\n",
        "        if wandb_log:\n",
        "            wandb.log({\n",
        "                \"iter\": iter_num,\n",
        "                \"train/loss\": losses['train'],\n",
        "                \"val/loss\": losses['val'],\n",
        "                \"lr\": lr,\n",
        "                \"mfu\": running_mfu*100, # convert to percentage\n",
        "            })\n",
        "        if losses['val'] < best_val_loss or always_save_checkpoint:\n",
        "            best_val_loss = losses['val']\n",
        "            if iter_num > 0:\n",
        "                checkpoint = {\n",
        "                    'model': raw_model.state_dict(),\n",
        "                    'optimizer': optimizer.state_dict(),\n",
        "                    'model_args': model_args,\n",
        "                    'iter_num': iter_num,\n",
        "                    'best_val_loss': best_val_loss,\n",
        "                    'config': config,\n",
        "                }\n",
        "                print(f\"saving checkpoint to {out_dir}\")\n",
        "                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))\n",
        "    if iter_num == 0 and eval_only:\n",
        "        break\n",
        "\n",
        "    # forward backward update, with optional gradient accumulation to simulate larger batch size\n",
        "    # and using the GradScaler if data type is float16\n",
        "    for micro_step in range(gradient_accumulation_steps):\n",
        "        if ddp:\n",
        "            # in DDP training we only need to sync gradients at the last micro step.\n",
        "            # the official way to do this is with model.no_sync() context manager, but\n",
        "            # I really dislike that this bloats the code and forces us to repeat code\n",
        "            # looking at the source of that context manager, it just toggles this variable\n",
        "            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)\n",
        "        with ctx:\n",
        "            logits, loss = model(X, Y)\n",
        "            loss = loss / gradient_accumulation_steps # scale the loss to account for gradient accumulation\n",
        "        # immediately async prefetch next batch while model is doing the forward pass on the GPU\n",
        "        X, Y = get_batch('train')\n",
        "        # backward pass, with gradient scaling if training in fp16\n",
        "        scaler.scale(loss).backward()\n",
        "    # clip the gradient\n",
        "    if grad_clip != 0.0:\n",
        "        scaler.unscale_(optimizer)\n",
        "        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)\n",
        "    # step the optimizer and scaler if training in fp16\n",
        "    scaler.step(optimizer)\n",
        "    scaler.update()\n",
        "    # flush the gradients as soon as we can, no need for this memory anymore\n",
        "    optimizer.zero_grad(set_to_none=True)\n",
        "\n",
        "    # timing and logging\n",
        "    t1 = time.time()\n",
        "    dt = t1 - t0\n",
        "    t0 = t1\n",
        "    if iter_num % log_interval == 0 and master_process:\n",
        "        # get loss as float. note: this is a CPU-GPU sync point\n",
        "        # scale up to undo the division above, approximating the true total loss (exact would have been a sum)\n",
        "        lossf = loss.item() * gradient_accumulation_steps\n",
        "        if local_iter_num >= 5: # let the training loop settle a bit\n",
        "            mfu = raw_model.estimate_mfu(batch_size * gradient_accumulation_steps, dt)\n",
        "            running_mfu = mfu if running_mfu == -1.0 else 0.9*running_mfu + 0.1*mfu\n",
        "        loss = Thread(target=estimate_loss)\n",
        "        print(f\"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%\")\n",
        "    iter_num += 1\n",
        "    local_iter_num += 1\n",
        "\n",
        "    # termination conditions\n",
        "    if iter_num > max_iters:\n",
        "        break\n",
        "\n",
        "\n",
        "def plot_losses(train_losses, val_losses):\n",
        "    plt.figure(figsize=(10, 5))\n",
        "    plt.plot(train_losses, label='Training Loss', color='blue')\n",
        "    plt.plot(val_losses, label='Validation Loss', color='red')\n",
        "    plt.xlabel('Epochs')\n",
        "    plt.ylabel('Loss')\n",
        "    plt.legend()\n",
        "    plt.title('Training & Validation Losses')\n",
        "    plt.show()\n",
        "\n",
        "plot_losses(train_losses, val_losses)\n",
        "\n",
        "\n",
        "if ddp:\n",
        "    destroy_process_group()\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "aScf-duTpktX"
      },
      "source": [
        "### Export the torch model to ONNX"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 5,
      "metadata": {
        "id": "mky5i8iNpktX",
        "scrolled": true,
        "outputId": "fd578328-7603-4af4-b8ae-7abc678d0dfb",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "number of parameters: 0.34M\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "<ipython-input-2-9148f21cb9e7>:305: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!\n",
            "  if idx.size(1) <= self.config.block_size\n",
            "<ipython-input-2-9148f21cb9e7>:255: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!\n",
            "  t <= self.config.block_size\n",
            "<ipython-input-2-9148f21cb9e7>:92: TracerWarning: Converting a tensor to a Python float might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!\n",
            "  att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))\n"
          ]
        }
      ],
      "source": [
        "# Sample text to be fed into the model for input.json creation. This can be any string.\n",
        "EXAMPLE_TEXT = 'Hello, my name is '\n",
        "\n",
        "# Paths for the export directory (out_dir), the pretrained model ckpt directory and the training data directory (containing meta.pkl)\n",
        "out_dir = 'model'\n",
        "train_data_dir = 'data'\n",
        "pretrained_dir = 'pretrained'\n",
        "\n",
        "# Joined paths\n",
        "checkpoint_path = os.path.join(pretrained_dir, \"ckpt.pt\")\n",
        "meta_path = os.path.join(train_data_dir, \"meta.pkl\")\n",
        "onnx_path = os.path.join(out_dir, \"network.onnx\")\n",
        "input_path = os.path.join(out_dir, \"input.json\")\n",
        "\n",
        "# Create out dir if not exists\n",
        "os.makedirs(out_dir, exist_ok=True)\n",
        "\n",
        "# Load the model\n",
        "checkpoint = torch.load(checkpoint_path)\n",
        "gptconf = GPTConfig(**checkpoint[\"model_args\"])\n",
        "model = GPT(gptconf)\n",
        "state_dict = checkpoint['model']\n",
        "\n",
        "# Remove unwanted prefix from some of the state dict params (from NanoGPT)\n",
        "unwanted_prefix = '_orig_mod.'\n",
        "for k,v in list(state_dict.items()):\n",
        "    if k.startswith(unwanted_prefix):\n",
        "        state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)\n",
        "\n",
        "# Export the model to ONNX and create input.json\n",
        "\n",
        "def export():\n",
        "    # Have tried using torch.compile, removing eval, moving eval below the load state dict... All with the same result.\n",
        "    model.eval()\n",
        "    # Load the trained state dict into the model\n",
        "    model.load_state_dict(state_dict, strict=True)\n",
        "\n",
        "    # Open the metadata pickle and load string to int and vice versa mappings so we can preview decoded text\n",
        "    with open(meta_path, 'rb') as f:\n",
        "        meta = pickle.load(f)\n",
        "    stoi, itos = meta['stoi'], meta['itos']\n",
        "    encode = lambda s: [stoi[c] for c in s]\n",
        "    decode = lambda l: ''.join([itos[i] for i in l])\n",
        "\n",
        "    # Convert sample text into tokens for passing into the model\n",
        "    tokenized_input = (torch.tensor(encode(EXAMPLE_TEXT), dtype=torch.long)[None, ...])\n",
        "\n",
        "    torch.onnx.export(\n",
        "                    model,\n",
        "                    tokenized_input,\n",
        "                    onnx_path,\n",
        "                    export_params=True,\n",
        "                    opset_version=14,\n",
        "                    do_constant_folding=True,\n",
        "                    input_names=['input'],\n",
        "                    output_names=['output'],\n",
        "                    dynamic_axes={'input': {0: 'batch_size'},'output': {0: 'batch_size'}},\n",
        "                )\n",
        "\n",
        "export()\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "PV6r15iwpktX"
      },
      "source": [
        "### Install EZKL CLI"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9QSEY1ygpktX"
      },
      "source": [
        "### Generate input.json"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {
        "id": "xLoXhC4apktX"
      },
      "outputs": [],
      "source": [
        "# Convert the input string into tokens using the encode method\n",
        "input_data = encode(EXAMPLE_TEXT)\n",
        "\n",
        "# Convert input data into tensor\n",
        "input_tensor = torch.tensor(input_data, dtype=torch.long)[None, ...]\n",
        "\n",
        "# Run the model\n",
        "output_tensor = model(input_tensor)\n",
        "\n",
        "# Convert output tensor to list\n",
        "output_data = output_tensor.tolist()[0]\n",
        "\n",
        "# Create a dictionary to hold the data\n",
        "data = {\"input_data\": [input_data], \"output_data\": [output_data]}\n",
        "\n",
        "# Define the path to the input.json file\n",
        "input_json_path = \"./model/input.json\"\n",
        "\n",
        "# Write the data to the input.json file\n",
        "with open(input_json_path, 'w') as f:\n",
        "    json.dump(data, f)\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "6ygqdFSOpktX"
      },
      "source": [
        "### Circuitize the model\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {
        "id": "7u1l91AWpktX",
        "outputId": "7bdc4b7d-a570-441a-d168-03a35b8e30e7",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Creating circuit for ./model/network.onnx...\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:30\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge Linear Algebra.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:30\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mGenSettings\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34margs\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcheck_mode\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mUNSAFE\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mdiv_rebasing\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: false,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34minput_scale\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 7,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34minput_visibility\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mPublic\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mlogrows\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 17,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mlookup_range\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m[\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m           -32768,\n",
            "\u001b[1;37m | \u001b[0m           32768\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m         ]\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_inner_cols\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 2,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34moutput_visibility\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mPublic\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mparam_scale\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 7,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mparam_visibility\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mFixed\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mscale_rebase_multiplier\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 1,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mtolerance\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m           \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mscale\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 1.0,\n",
            "\u001b[1;37m | \u001b[0m           \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mval\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 0.0\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m         }\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m         \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mvariables\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m[\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m           \u001b[1m[\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m             \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mbatch_size\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m             1\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m           ]\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m         ]\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m       }\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mmodel\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.onnx\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msettings_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/settings.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:30\u001b[0m, ezkl::graph::model] - \u001b[1;37mset batch_size to 1\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:31\u001b[0m, ezkl::graph::model] - \u001b[1;37m\u001b[34mmodel has\u001b[0m\u001b[1;37m \u001b[34m2\u001b[0m\u001b[1;37m \u001b[34minstances\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:31\u001b[0m, ezkl::graph::model] - \u001b[1;37mcalculating num of constraints using dummy model layout...\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:34\u001b[0m, ezkl::graph::model] - \u001b[1;37m\u001b[34mmodel uses\u001b[0m\u001b[1;37m \u001b[34m11823993\u001b[0m\u001b[1;37m \u001b[34mrows\u001b[0m\u001b[1;37m (coord=\u001b[33m23647534\u001b[0m\u001b[1;37m, constants=\u001b[31m20259673\u001b[0m\u001b[1;37m)\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:34\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:34\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for Layers.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:34\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mCalibrateSettings\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mdata\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/input.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mdiv_rebasing\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mlookup_safety_margin\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 2,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mmax_logrows\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mmodel\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.onnx\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mscale_rebase_multiplier\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m[\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m         1,\n",
            "\u001b[1;37m | \u001b[0m         2,\n",
            "\u001b[1;37m | \u001b[0m         10\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m       ]\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mscales\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msettings_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/settings.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mtarget\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mAccuracy\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:35\u001b[0m, ezkl::execute] - \u001b[1;37mnum of calibration batches: 1\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:35\u001b[0m, ezkl::execute] - \u001b[1;37mrunning onnx predictions...\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:35\u001b[0m, ezkl::graph::model] - \u001b[1;37mset batch_size to 1\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:35\u001b[0m, tract_linalg::x86_64_fma] - \u001b[1;37mqmmm_i32: x86_64/avx2 activated\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:43:35\u001b[0m, tract_linalg::x86_64_fma] - \u001b[1;37mmmm_f32, mmv_f32, sigmoid_f32, tanh_f32: x86_64/fma activated\u001b[0m\n",
            "\u001b[2K\u001b[1A[00:03:07] \u001b[36m########################################\u001b[34m\u001b[0m\u001b[0m      24/24      Calibration Done.\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:45\u001b[0m, ezkl::execute] - \u001b[1;33m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m  <------------- Numerical Fidelity Report (input_scale: 0, param_scale: 13, scale_input_multiplier: 1) ------------->\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m +------------+--------------+-----------+-----------+----------------+------------------+---------------+---------------+--------------------+--------------------+------------------------+\n",
            "\u001b[1;37m | \u001b[0m | mean_error | median_error | max_error | min_error | mean_abs_error | median_abs_error | max_abs_error | min_abs_error | mean_squared_error | mean_percent_error | mean_abs_percent_error |\n",
            "\u001b[1;37m | \u001b[0m +------------+--------------+-----------+-----------+----------------+------------------+---------------+---------------+--------------------+--------------------+------------------------+\n",
            "\u001b[1;37m | \u001b[0m | 0          | 0            | 0         | 0         | 0              | 0                | 0             | 0             | 0                  | 0                  | 0                      |\n",
            "\u001b[1;37m | \u001b[0m +------------+--------------+-----------+-----------+----------------+------------------+---------------+---------------+--------------------+--------------------+------------------------+\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:45\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:45\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for Layers.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:45\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mGetSrs\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcheck\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mSAFE\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mlogrows\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msettings_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/settings.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msrs_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/kzg.srs\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:45\u001b[0m, ezkl::execute] - \u001b[1;37mSRS already exists at that path\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:55\u001b[0m, ezkl::execute] - \u001b[1;37mSRS hash: dec49a69893fbcd66cd06296b2d936a6aceb431c130b2e52675fe4274b504f57\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:55\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:55\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for Liberty.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:55\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mSetup\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcompiled_circuit\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.ezkl\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcompress_selectors\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: false,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mpk_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/pk.key\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msrs_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/kzg.srs\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mvk_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/vk.key\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mwitness\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:46:55\u001b[0m, ezkl::pfsys::srs] - \u001b[1;37mloading srs from \"./model/kzg.srs\"\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:01\u001b[0m, ezkl::execute] - \u001b[1;37mdownsizing params to 17 logrows\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::graph::vars] - \u001b[1;37mnumber of blinding factors: 5\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 154 fixed columns\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::graph::model] - \u001b[1;37mconfiguring model\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:33\u001b[0m, ezkl::graph] - \u001b[1;37mcircuit size: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_advice_columns\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 546,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_challenges\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 0,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_fixed\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 168,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_instances\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 1,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_selectors\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 4004\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:35\u001b[0m, ezkl::circuit::modules::planner] - \u001b[1;33mspawning module 2\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:47:35\u001b[0m, ezkl::graph::model] - \u001b[1;37mmodel layout...\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:54:36\u001b[0m, ezkl::graph::model] - \u001b[1;37m\u001b[34mmodel uses\u001b[0m\u001b[1;37m \u001b[34m11803919\u001b[0m\u001b[1;37m \u001b[34mrows\u001b[0m\u001b[1;37m (coord=\u001b[33m23607838\u001b[0m\u001b[1;37m, constants=\u001b[31m18904174\u001b[0m\u001b[1;37m)\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::pfsys] - \u001b[1;37mVK took 512.225\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::graph::vars] - \u001b[1;37mnumber of blinding factors: 5\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 90 advice blocks\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::tensor::var] - \u001b[1;33musing column duplication for 154 fixed columns\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::graph::model] - \u001b[1;37mconfiguring model\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::graph] - \u001b[1;37mcircuit size: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_advice_columns\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 546,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_challenges\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 0,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_fixed\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 168,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_instances\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 1,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_selectors\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 4004\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;33mW\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::circuit::modules::planner] - \u001b[1;33mspawning module 2\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 20:56:05\u001b[0m, ezkl::graph::model] - \u001b[1;37mmodel layout...\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:03:07\u001b[0m, ezkl::graph::model] - \u001b[1;37m\u001b[34mmodel uses\u001b[0m\u001b[1;37m \u001b[34m11803919\u001b[0m\u001b[1;37m \u001b[34mrows\u001b[0m\u001b[1;37m (coord=\u001b[33m23607838\u001b[0m\u001b[1;37m, constants=\u001b[31m18904174\u001b[0m\u001b[1;37m)\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:09:56\u001b[0m, ezkl::pfsys] - \u001b[1;37mPK took 831.481\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:09:56\u001b[0m, ezkl::pfsys] - \u001b[1;37msaving verification key 💾\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:09:56\u001b[0m, ezkl::pfsys] - \u001b[1;37msaving proving key 💾\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:15:10\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:15:10\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for Neural Networks.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:15:10\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mCompileCircuit\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcompiled_circuit\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.ezkl\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mmodel\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.onnx\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msettings_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/settings.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:15:10\u001b[0m, ezkl::graph::model] - \u001b[1;37mset batch_size to 1\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:15:11\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "Time taken to circuitize model:  1900.9179220199585\n"
          ]
        }
      ],
      "source": [
        "# Define the base directory\n",
        "import os\n",
        "model_dir = \"./model/\"\n",
        "\n",
        "# Constructing paths using os.path.join\n",
        "onnx_path = os.path.join(model_dir, \"network.onnx\")\n",
        "input_path = os.path.join(model_dir, \"input.json\")\n",
        "settings_path = os.path.join(model_dir, \"settings.json\")\n",
        "srs_path = os.path.join(model_dir, \"kzg.srs\")\n",
        "ezkl_path = os.path.join(model_dir, \"network.ezkl\")\n",
        "pk_path = os.path.join(model_dir, \"pk.key\")\n",
        "vk_path = os.path.join(model_dir, \"vk.key\")\n",
        "witness_path = os.path.join(model_dir, \"witness.json\")\n",
        "proof_path = os.path.join(model_dir, \"proof.proof\")\n",
        "sol_path = os.path.join(model_dir, \"verif.sol\")\n",
        "abi_path = os.path.join(model_dir, \"verif.abi\")\n",
        "ezkl_binary = '/' + os.path.join('root', '.ezkl', 'ezkl')\n",
        "\n",
        "# Printing initial message\n",
        "print(f\"Creating circuit for {onnx_path}...\")\n",
        "\n",
        "# Executing the commands\n",
        "import time\n",
        "start = time.time()\n",
        "!{ezkl_binary} gen-settings -M {onnx_path} --param-visibility fixed --input-visibility public --output-visibility public -O {settings_path}\n",
        "!{ezkl_binary} calibrate-settings -M {onnx_path} -D {input_path} -O {settings_path} --target resources\n",
        "!{ezkl_binary} get-srs -S {settings_path} --srs-path {srs_path}\n",
        "!{ezkl_binary} setup -M {ezkl_path} --vk-path {vk_path} --pk-path {pk_path} --srs-path {srs_path}\n",
        "!{ezkl_binary} compile-circuit -M {onnx_path} -S {settings_path} --compiled-circuit {ezkl_path}\n",
        "end = time.time()\n",
        "print(\"Time taken to circuitize model: \", end - start)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "JLo-BlfspktX"
      },
      "source": [
        "### Reproduction of the issue\n",
        "\n",
        "When attempting to generate a proof for this model, an error is thrown."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {
        "id": "NUW1u_v0pktX",
        "scrolled": true,
        "outputId": "009bdb35-896e-45ab-ec3a-cf6c98a781a2",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        }
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Running\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:42\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for the Laconic.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:42\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mGenWitness\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcompiled_circuit\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.ezkl\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mdata\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/input.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34moutput\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/witness.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msrs_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mvk_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: null\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:42\u001b[0m, ezkl::graph] - \u001b[1;37minput scales: [0]\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:46\u001b[0m, ezkl] - \u001b[1;37msucceeded\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:46\u001b[0m, ezkl] - \u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██╗  ██╗██╗\n",
            "\u001b[1;37m | \u001b[0m         ██╔════╝╚══███╔╝██║ ██╔╝██║\n",
            "\u001b[1;37m | \u001b[0m         █████╗    ███╔╝ █████╔╝ ██║\n",
            "\u001b[1;37m | \u001b[0m         ██╔══╝   ███╔╝  ██╔═██╗ ██║\n",
            "\u001b[1;37m | \u001b[0m         ███████╗███████╗██║  ██╗███████╗\n",
            "\u001b[1;37m | \u001b[0m         ╚══════╝╚══════╝╚═╝  ╚═╝╚══════╝\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m         Easy Zero Knowledge for the Lyrical.\n",
            "\u001b[1;37m | \u001b[0m         -----------------------------------------------------------\n",
            "\u001b[1;37m | \u001b[0m \n",
            "\u001b[1;37m | \u001b[0m         \u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:46\u001b[0m, ezkl] - \u001b[1;37mcommand: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcommand\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m     \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mProve\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcheck_mode\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mSAFE\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mcompiled_circuit\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/network.ezkl\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mpk_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/pk.key\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mproof_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/proof.proof\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mproof_type\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32mSingle\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34msrs_path\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/kzg.srs\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m,\n",
            "\u001b[1;37m | \u001b[0m       \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mwitness\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: \u001b[32m\"\u001b[0m\u001b[1;37m\u001b[32m./model/witness.json\u001b[0m\u001b[1;37m\u001b[32m\"\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m     }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m   }\u001b[0m\u001b[1;37m\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:46\u001b[0m, ezkl::pfsys::srs] - \u001b[1;37mloading srs from \"./model/kzg.srs\"\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl::execute] - \u001b[1;37mdownsizing params to 25 logrows\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl::pfsys] - \u001b[1;37mloading proving key from \"./model/pk.key\"\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl::graph::vars] - \u001b[1;37mnumber of blinding factors: 5\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl::graph::model] - \u001b[1;37mconfiguring model\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;34m*\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl::graph] - \u001b[1;37mcircuit size: \n",
            "\u001b[1;37m | \u001b[0m  \u001b[1m{\u001b[0m\u001b[1;37m\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_advice_columns\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 6,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_challenges\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 0,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_fixed\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 14,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_instances\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 1,\n",
            "\u001b[1;37m | \u001b[0m   \u001b[1;34m\"\u001b[0m\u001b[1;37m\u001b[1;34mnum_selectors\u001b[0m\u001b[1;37m\u001b[1;34m\"\u001b[0m\u001b[1;37m: 44\u001b[1m\n",
            "\u001b[1;37m | \u001b[0m }\u001b[0m\u001b[1;37m\u001b[0m\n",
            "\u001b[1;34m[\u001b[0m\u001b[1;31mE\u001b[0m\u001b[1;34m]\u001b[0m [\u001b[95m2024-02-01 21:32:52\u001b[0m, ezkl] - \u001b[1;31mfailed: input number is not less than field modulus\u001b[0m\n",
            "Error: Custom { kind: InvalidData, error: \"input number is not less than field modulus\" }\n",
            "Witness + Proof took10.627288579940796\n"
          ]
        },
        {
          "output_type": "error",
          "ename": "ValueError",
          "evalue": "invalid literal for int() with base 16: 'E 2024-02-01 21:32:52'",
          "traceback": [
            "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
            "\u001b[0;31mValueError\u001b[0m                                Traceback (most recent call last)",
            "\u001b[0;32m<ipython-input-12-131d3f205002>\u001b[0m in \u001b[0;36m<cell line: 21>\u001b[0;34m()\u001b[0m\n\u001b[1;32m     19\u001b[0m ]\n\u001b[1;32m     20\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 21\u001b[0;31m instances = [\n\u001b[0m\u001b[1;32m     22\u001b[0m     \u001b[0mint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mhex_string\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstrip\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m16\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     23\u001b[0m     \u001b[0;32mfor\u001b[0m \u001b[0mhex_string\u001b[0m \u001b[0;32min\u001b[0m \u001b[0minstances_string\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mreplace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"[\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mreplace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"]\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msplit\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\",\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
            "\u001b[0;32m<ipython-input-12-131d3f205002>\u001b[0m in \u001b[0;36m<listcomp>\u001b[0;34m(.0)\u001b[0m\n\u001b[1;32m     20\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     21\u001b[0m instances = [\n\u001b[0;32m---> 22\u001b[0;31m     \u001b[0mint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mhex_string\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstrip\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m16\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     23\u001b[0m     \u001b[0;32mfor\u001b[0m \u001b[0mhex_string\u001b[0m \u001b[0;32min\u001b[0m \u001b[0minstances_string\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mreplace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"[\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mreplace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"]\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msplit\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\",\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     24\u001b[0m ]\n",
            "\u001b[0;31mValueError\u001b[0m: invalid literal for int() with base 16: 'E 2024-02-01 21:32:52'"
          ]
        }
      ],
      "source": [
        "print(\"Running\");\n",
        "start = time.time()\n",
        "!{ezkl_binary} gen-witness -D {input_path} -M {ezkl_path} -O {witness_path}\n",
        "!{ezkl_binary} prove -M {ezkl_path} -W {witness_path} --pk-path {pk_path} --proof-path {proof_path} --srs-path {srs_path}\n",
        "end = time.time()\n",
        "print(\"Witness + Proof took\" + str(end - start))\n",
        "\n",
        "proof_in_hex = subprocess.run(\n",
        "    [        ezkl_binary,\"print-proof-hex\", \"--proof-path\", proof_path],\n",
        "    text=True,\n",
        "    stdout=subprocess.PIPE,\n",
        "    stderr=subprocess.PIPE,\n",
        ").stdout\n",
        "proof_in_hex = remove_non_ascii(proof_in_hex)\n",
        "\n",
        "split_by_command = proof_in_hex.split(\"|  }\")[1]\n",
        "instances_string = split_by_command.split(\"[*] [0s, ezkl::execute] - \")[0].split(\"\\n\")[\n",
        "    1\n",
        "]\n",
        "\n",
        "instances = [\n",
        "    int(hex_string.strip(), 16)\n",
        "    for hex_string in instances_string.replace(\"[\", \"\").replace(\"]\", \"\").split(\",\")\n",
        "]\n",
        "\n",
        "\n",
        "print(\"Instances from the circuit:\", instances)\n",
        "print(\"Decoded instances from the circuit: \\n\", \"\".join(decode([i]) for i in instances))\n",
        "print(\"Output tokens from PyTorch: \", output_data)\n",
        "print(\"Decoded output from PyTorch: \\n\", \"\".join(decode([i]) for i in output_data))\n",
        "\n",
        "\n",
        "session = ort.InferenceSession(onnx_path, providers=[\"CPUExecutionProvider\"])\n",
        "results = session.run(\n",
        "    None,\n",
        "    {k.name: [v.numpy().tolist()] for k, v in zip(session.get_inputs(), input_tensor)},\n",
        ")\n",
        "print(\"Output tokens from ONNX\", results)\n",
        "print(\"Decoded output from ONNX: \\n\", decode(results[0][0]))\n"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "provenance": [],
      "include_colab_link": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.8.6"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}